Cue-Reflex-4B v0.2v1.7 · 10 s rule
Data: JevBench v1.8.18, published 2026-11-20 · method v1.7 (10 s rule), measured under the rule
Capability 64.7 · rank #113 of 251 on the open-weights board
Last measured: 2026-11-17 · measurement release: v1.8.15
Apache-2.0 (repo LICENSE and inference/LICENSE at a36bd3ec; model card apache-2.0); base google/gemma-4-E4B-it Apache-2.0
intelligence
50.9
calibration
78.4
speed
89.2
cost
87.4
Composite (secondary)
72.7 · Composite rank #55
Cost and measurement conditions
$0.0026 per 1,000 decisions (estimated)
GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, USD 1.5235/hour; average load 70%. Architecture: small gemma, 4B active parameters; throughput 80585.75 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.00262241966066776/1,000; old list-price estimate USD 0.00699114842175957/1,000 (deepinfra:google/gemma-4-E4B-it at USD 0.02/M input and USD 0.10/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00262241966066776/1,000. Common 1,489-item basis: 520,491 input and zero output tokens. Estimated, not a bill.
Measured p50 latency: 0.095 s (on-GPU, no network, loopback; not directly comparable with API latency, which we time end to end from our measuring host, network included)
Ranking and cap use: 0.34 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.095 s on-GPU, no network (timed on our GPU over loopback). Not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and latency cap use the same estimate for every self-hosted row: 2 × measured + 150 ms = 0.34 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
Evaluator-owned Verda GPU VM (RTX6000), offline read-only container (--network none); author inference/serve.py at a36bd3ec through a first-party launcher; BF16 Gemma-4-E4B + LoRA with a frozen answer-letter readout, one forward pass, zero output tokens. Pinned inference_config: sandwich prompt format, per-kind temperatures 1.118/1.032/1.227 and the disclosed yes/no confidence rule; limits 8192 prompt tokens and 16 options, no truncation, HTTP 422 above limits. PyTorch 2.9.1 cu128, transformers 5.20.1, peft 0.21.2 (adapter metadata records 0.21.3), Python 3.11.
Measured latency over all decisions: p50 0.095 s · p95 0.103 s
— decisions over 10 s.
A decision that takes longer than 10 seconds (measured time) counts as a wrong answer and as zero calibration for that item: each type's calibration is multiplied by the share of its decisions answered within 10 s. The limit uses the measured time (raw). The displayed adjusted time for self-hosted rows (2 × + 0.15 s) is an assumption and is not used for the limit.
Breakdowns (not part of the score)
Chance-corrected competence (0 = chance, 100 = perfect) by category on open+sealed. Raw values, not equated; the same cells feed the compare view.
Capability by subject topic
Use cases (TypeSafe categories)
| Model routing | 72.0 | n=148 |
|---|---|---|
| Legal & compliance | 52.3 | n=56 |
| Customer support | 48.5 | n=66 |
| E-commerce | 57.8 | n=62 |
| Insurance claims | 37.1 | n=62 |
| Risk assessment | 60.1 | n=48 |
| Financial crime | 46.1 | n=63 |
| Feature extraction | 58.2 | n=72 |
| Lead generation | 69.1 | n=75 |
| Recruiting | 75.4 | n=65 |
| Knowledge graphs | 59.1 | n=67 |
| LLM guardrails | 78.5 | n=71 |
| Moderation | 78.1 | n=59 |
| Code linting | 43.9 | n=69 |
| Search & retrieval | 85.3 | n=67 |
| Science | 34.9 | n=42 |
| Advertising | 60.4 | n=48 |
| Gaming | 57.1 | n=58 |
| Demand forecasting | 8.9 | n=66 |
Cells need at least 30 answered items; † fewer than 30 (indicative only; drawn as hollow, lighter points on the compare radars).
Model and serving disclosure
Paid public evaluation. Measured on the 27 Oct draw on an evaluator-owned rented GPU, offline with the author's serving stack at the pinned commit. Failed or unanswered requests count as wrong answers. Gemma-4-E4B finetune with a frozen answer-letter readout in one forward pass and zero output generation. The serving configuration applies a yes/no confidence rule: yes-probability in (0.20, 0.50) becomes 0.199, and [0.50, 0.80) becomes 0.801. This rule is part of the measured model. Over-limit items are refused with HTTP 422 and scored wrong.